feat(ENG-CUDAGRAPH-DEDUP): fold captures that share a topology onto one graph executable - #1178

Merged

localai-bot merged 14 commits into

mainfrom

row/ENG-CUDAGRAPH-DEDUP

Aug 18, 2026

Commits on Aug 18, 2026

spec(ENG-CUDAGRAPH-DEDUP): one graph executable per topology, not one per bucket
mudler
committed
6639f41
feat(ENG-CUDAGRAPH-DEDUP): fold captures that share a topology onto one graph executable
mudler
committed
c1ca9ac
chore(ENG-CUDAGRAPH-DEDUP): re-pin RUNNABLE_BASELINE for the row that just entered it
mudler
committed
ff82379
merge: origin/main into row/ENG-CUDAGRAPH-DEDUP -- no keyed record overlapped
mudler
committed
8d381f1
test(ENG-CUDAGRAPH-DEDUP): pin the new RUNNABLE_BASELINE entry with its own mutation
mudler
committed
89635e4
merge: origin/main into row/ENG-CUDAGRAPH-DEDUP -- no keyed record overlapped
mudler
committed
6fd2e59
fix(ENG-CUDAGRAPH-DEDUP): bind both shapes of cudaGraphGetEdges, which changed at CUDA 13
mudler
committed
e4ce557
merge(ENG-CUDAGRAPH-DEDUP): take both engine-matrix row sets, and keep #1179's corrected driver count
mudler
committed
469ac59
fix(ENG-CUDAGRAPH-DEDUP): a failed instantiate now throws where the capture happened, not at the first replay
mudler
committed
a0b8ff3
style(ENG-CUDAGRAPH-DEDUP): reflow the header paragraph the driver-count correction broke
mudler
committed
5791c8b
test(ENG-CUDAGRAPH-DEDUP): gate the loud-failure claim the transitivity note now rests on
mudler
committed
3ed6253
merge(ENG-CUDAGRAPH-DEDUP): take origin/main, and re-derive the three counts both sides moved
mudler
committed
14061e2
fix(ENG-CUDAGRAPH-DEDUP): a file allowed to see the runtime FAIL must consume the latched error, and this one never did (#1184, #1162)
mudler
committed
72de552
merge: origin/main into row/ENG-CUDAGRAPH-DEDUP
mudler
committed
1a0117a

Navigation Menu

feat(ENG-CUDAGRAPH-DEDUP): fold captures that share a topology onto one graph executable - #1178

feat(ENG-CUDAGRAPH-DEDUP): fold captures that share a topology onto one graph executable#1178
localai-bot merged 14 commits into
mainfrom
row/ENG-CUDAGRAPH-DEDUP

Commits on Aug 18, 2026

spec(ENG-CUDAGRAPH-DEDUP): one graph executable per topology, not one per bucket

feat(ENG-CUDAGRAPH-DEDUP): fold captures that share a topology onto one graph executable

chore(ENG-CUDAGRAPH-DEDUP): re-pin RUNNABLE_BASELINE for the row that just entered it

merge: origin/main into row/ENG-CUDAGRAPH-DEDUP -- no keyed record overlapped

test(ENG-CUDAGRAPH-DEDUP): pin the new RUNNABLE_BASELINE entry with its own mutation

merge: origin/main into row/ENG-CUDAGRAPH-DEDUP -- no keyed record overlapped

fix(ENG-CUDAGRAPH-DEDUP): bind both shapes of cudaGraphGetEdges, which changed at CUDA 13

merge(ENG-CUDAGRAPH-DEDUP): take both engine-matrix row sets, and keep #1179's corrected driver count

fix(ENG-CUDAGRAPH-DEDUP): a failed instantiate now throws where the capture happened, not at the first replay

style(ENG-CUDAGRAPH-DEDUP): reflow the header paragraph the driver-count correction broke

test(ENG-CUDAGRAPH-DEDUP): gate the loud-failure claim the transitivity note now rests on

merge(ENG-CUDAGRAPH-DEDUP): take origin/main, and re-derive the three counts both sides moved

fix(ENG-CUDAGRAPH-DEDUP): a file allowed to see the runtime FAIL must consume the latched error, and this one never did (#1184, #1162)

merge: origin/main into row/ENG-CUDAGRAPH-DEDUP